The federal government labeled Anthropic a "supply chain risk" in February 2025. The Department of Defense moved to cut off Anthropic and force its vendors to follow suit. Six months later, the NSA is running Mythos Preview, the DoD is using it widely, and the restriction exists mostly on paper.
The Gap Between Policy and Practice
This isn't a new pattern. The TEXXR signal data shows federal agencies have been quietly expanding their use of frontier AI models despite increasingly restrictive rhetoric. The designation was meant to signal toughness on Chinese AI ties — but it didn't account for the fact that American agencies actually need these models.
The NSA's use of Mythos makes sense from a capability standpoint. Mythos is Anthropic's most powerful model, designed for long-context reasoning and complex analysis. It's the kind of tool that helps with signal processing, threat assessment, and policy drafting. The alternative — not having the best AI available — is a harder sell internally than explaining away a procurement exception.
When the gap between what's restricted and what's useful gets large enough, agencies find ways to bridge it.
What This Means for the Industry
For AI companies trying to sell to the federal government, this episode reveals something important: the procurement restrictions are negotiable when capability matters. Anthropic wasn't removed from government systems — it was quietly re-added through the back door when the DoD realized what it was giving up.
This also suggests the $800B in VC offers flowing to AI labs aren't purely speculative. The government's own use of these models — despite official discouragement — validates the technical moat. If the most security-conscious agencies in the country can't replace Anthropic's output with alternatives, that's a market signal worth listening to.
The practical takeaway for enterprise AI buyers: expect government procurement patterns to be less restrictive in practice than the rhetoric suggests. Watch what agencies actually do, not what they say in press releases.